Invotet logo

RVPU · configurable robotics vision

Robotics Vision, Deployed in a Day

The RVPU — Robotics Vision Processing Unit — is a set of perception functions you wire cameras into. Pick the cameras your SWaP, cost, and supply chain allow, mount them in the geometry your robot needs, and get dense 3D, detected and tracked objects, and thermal–RGB matches back over ROS 2. It is deliberately not a generic AI accelerator: perception as a unit, not a model-compile path and a pipeline you have to build yourself.

Become a design partnerSee the RVPU
TerraBot X — RVPU on an AMD Kria system-on-moduleTop-down illustration of the TerraBot X RVPU vision appliance on an AMD Kria system-on-modulePIPELINELIBRARYRVPUVISIONINVOTETROS 2TRACE 8192CAMERAINMIPI CSI-2ROS 23D · TRACKSOUTAMD KRIA SOM
RVPU · ground robotics · Kria KV260

TerraBot X

Your cameras · dense 3D, detect & track · ROS 2 native · on-platform

AMD Kria KV260RVPU host
ROS 2 · 3D + tracksPerception out
Your rigUSB / CSI / MIPI / LVDS
Self-calibratingDrift-compensating
8,192Trace buffer entries
No model codeApp config, not AI
Design partner program

Pre-compiled vision pipelines · in-mission bitstream switching · 8,192-entry cycle-accurate hardware trace buffer

  • Demo target

    30+ Hz

    Depth Anything V2 · Kria KV260 · preliminary

  • Cameras

    Yours

    any baseline, any layout

  • Deploy

    1 day

    not a six-month AI program

  • Trace

    8,192

    cycle-accurate trace entries

Designed for

Qualification on roadmap · reports to design partners under NDA

MIL-STD-810H

Vibration + shock — designed for, qualification on roadmap

−40 to +85 °C

Target operating range

Ruggedized

Ingress-protected enclosure (target)

ROS 2 · GStreamer

Native autonomy integration

Trace buffer

8,192-entry cycle-accurate · forensic observability

Secure boot

Hardware root of trust · per-module attestation

Designed for

The buyers who can't ship on commercial silicon

The problem

Design freedom, or a vendor's camera kit. Pick one.

Off-the-shelf 3D and stereo platforms force a specific vendor camera kit, fixing your baseline, placement, and field of view. The alternative — designing your own stereo rig, or a six-camera 360° setup — is custom engineering work most teams cannot afford to keep reinventing. The RVPU takes the third path: bring cameras of the focus and FoV you want, place them how your platform demands, connect them to the unit, and get 3D and higher-level perception back.

RVPU · ground robotics · Kria KV260

TerraBot X

RVPU perception unit for ground robotics — dense stereo 3D from cameras you choose, automatic self-calibration, 3D detection and tracking, and thermal–RGB cross-modality matching on an AMD Kria KV260.

Operating range
−40 to +85 °C (target)
Perception
Ground robotics · dense 3D, self-calibration, detection & tracking, thermal–RGB matching
Design partner program
RVPU detail

Why the RVPU

Deploy robotics vision in a day — not a six-month AI program.

How the RVPU works

  • Perception as a unit

    The RVPU is not a generic AI accelerator. It is a curated set of robotics perception functions — dense 3D, detection, segmentation, tracking, cross-modality matching — exposed as configured pipelines. You wire cameras in and get useful outputs back, without building a perception stack first.

  • Your cameras, your geometry

    Off-the-shelf 3D platforms force a vendor camera kit and fix your baseline, placement, and field of view. The RVPU does the opposite: pick the cameras your SWaP, cost, and supply chain allow, place them how your platform needs, and the unit works with that rig.

  • Calibrates itself

    After one calibration against reference frames, the unit re-calibrates automatically to compensate for drift. A rig that shifts under vibration or thermal cycling keeps producing depth you can trust — no field re-calibration procedure.

  • On-platform, link-independent

    Perception runs on the robot. No tethered GPU, no cloud round-trip, no degraded behaviour when the uplink drops. What we optimize for is camera-to-output latency, frames per watt, and multi-camera throughput — not raw TOPS.

  • Sovereign & resilient

    Defense-grade and built for sovereign supply — deployable under any country or region’s sovereignty requirements, with no dependence on a foreign cloud or data path. Sovereign and OEM programs can license the SDK and IP to keep the perception stack under their own control.

  • Secure by design, forensic by default

    Hardware root of trust, signed firmware, and per-module attestation keep the device tamper-evident. An 8,192-entry cycle-accurate hardware trace buffer gives forensic-grade observability for mission review and ROE compliance.

Invotet SDK

Describe your rig. Run the pipeline.

A Python SDK for configuring the RVPU — describe your cameras and their geometry, pick the perception functions you need, and stream 3D positions, depth maps, and tracks into ROS 2. It ingests PyTorch, ONNX, and HuggingFace models when a pipeline needs a custom detector, with no CUDA in the loop. App config, not model code.

Explore the SDKBecome a design partner

  • Framework

    PyTorch

    Bring a custom detector: trace or torch.export models fold into a pipeline with no rewrite.

  • Framework

    ONNX

    Standards-based interchange — any ONNX-exported model can back a pipeline stage.

  • Framework

    HuggingFace

    Vision checkpoints load through a one-line loader when a pipeline is customised.

For developers

Real entry points, not a wall of marketing

Bring your cameras and your autonomy stack.

The RVPU is available through our design-partner program. Tell us the rig — how many cameras, what baseline, which modalities — plus the perception you need and the integration target. We will set up an evaluation and share the right documentation, including qualification reports under NDA as they complete.

Become a design partner